In this post, I will try to compare and contrast Julia, R, and Python via a simple maximum likelihood optimization problem whose motivation comes from the credit risk domain and is discussed in more detail inthis post

#### Side note - the author's experience level at the time of writing

Language | Level of experience |
---|---|

R | 9 years |

Julia | 6 months |

Python | beginner |

For such a simple optimization problem, R, Julia, and Python/SciPy will **all do a competent job** , so there is no clear winner. However as noted by Julia discourse member ChrisRackauckas

if you wish to solve only a single optimization problem and that problem takes <10 seconds, then Julia’s long initial compilation is something you want to avoid. For long enough problems, or for solving multiple optimization problems, the compilation time is not noticeable.

## The optimization problem

Given observations

Q

1

,

Q

2

,

.

.

.

,

Q

n

Q_1,, Q_2,, ...,, Q_n

Q 1 , Q 2 , . . . , Q n , we aim to find paramters

μ

mu

and

σ

sigma

that optimize this likelihood function

L

=

∏

(

ϕ

(

Q

i

,

μ

,

σ

)

Φ

(

max

Q

t

,

μ

,

σ

)

)

L = prodleft(frac{phi(Q_i,mu,sigma)}{Phi(max Q_t,mu,sigma)}right)

L = ∏ ( Φ ( max Q t , μ , σ ) ϕ ( Q i , μ , σ ) )

often we try to optimize the log-likelihood instead

log

L

=

l

=

(

∑

i

ϕ

(

Q

i

,

μ

,

σ

)

)

−

n

Φ

(

max

Q

t

,

μ

,

σ

)

log L = l = left(sum_i phi(Q_i,mu,sigma)right) - nPhi(max Q_t,mu,sigma)

lo g L = l = ( i ∑ ϕ ( Q i , μ , σ ) ) − n Φ ( max Q t , μ , σ )

side note - a more efficient numerical solution exists |
---|

This Ergashev et al., 2016 paper (behind a paywall) derived a necessary condition for when a solution exists and derived an equation in one parameter which when solved (using simple uniroot techniques) can recover both μ mu and σ sigma . From my testing, applying Ergashev's formula yields about 50x speed up to the R solution. However, in this blogpost, I aim to compare and contrast the optimization function in Julia vs. R vs. Python and hence I have chosen not to implement Ergashev's methods. |

## Julia solution

Below is my Julia implementation using `Optim.jl`

In Julia, one can use **symbols** in variable names, so I have used

μ

σ

musigma

as a variable name. Julia also has a popular package called JuMP.jl for optimization problems. However, the Optim.jl package is more than adequate for such a simple problem, and I will only look at JuMP.jl in the future when dealing with more advanced optimization problems.

I have noted the time it takes to perform the first run of the optimization problem is 7.5 seconds. This is considerably slower than both R and Python. In regards to this, Julia Discourse member ChrisRackauckas have pointed out

If you want to solve a problem that takes 3000 seconds to solve, the first time in a Julia session will make it take 3007 seconds making the total runtime more relevant to the total time than compilation. If you want to solve many 100 10 second optimization problems, the first will take 17 seconds, and subsequent calls will not have the compilation and will take roughly 10 seconds, making the total runtime 1007 seconds and thus the individual problem each is more relevant than the compilation time. If you want to solve a single 5-second optimization problem, it’ll be 12 seconds. Thus the compilation time issue can be annoying for interactive use when you want to solve a single problem but doesn’t effect performance-sensitive uses.

He also notes that

Julia v0.7 does have an interpreter though which promises to reduce compilation time in exchange for less optimizations and this may be an interesting middle ground for the specific interactive use case. Additionally, tools like PackageCompiler.jl may become more common, limiting the “startup time” of package code that is noticed here.

In the below, I have hard-coded the values of `Q_t`

that I would like to use in the MLE estimation.

using Distributions, Optim # hard coded dataobservations odr=[0.10,0.20,0.15,0.22,0.15,0.10,0.08,0.09,0.12] Q_t = quantile.(Normal(0,1), odr) # return a function that accepts `[mu, sigma]` as parameter function neglik_tn(Q_t) maxx = maximum(Q_t) f(μσ) = -sum(logpdf.(Truncated(Normal(μσ[1],μσ[2]), -Inf, maxx), Q_t)) f end neglikfn = neglik_tn(Q_t) # optimize! # start searching @time res = optimize(neglikfn, [mean(Q_t), std(Q_t)]) # 7.5 seconds @time res = optimize(neglikfn, [mean(Q_t), std(Q_t)]) # 0.000137 seconds # the mu and sigma estimates Optim.minimizer(res) # [-1.0733250637041452,0.2537450497038758] # or # use `fieldnames(res)` to see the list of field names that can be referenced via . (dot) res.minimizer # [-1.0733250637041452,0.2537450497038758]

which gave me the following output which is by far the most descriptive out of the Julia/R/Python-verse.

Results of Optimization Algorithm * Algorithm: Nelder-Mead * Starting Point: [-1.1300664159893685,0.22269345618402703] * Minimizer: [-1.0733250637041452,0.2537450497038758] * Minimum: -1.893080e+00 * Iterations: 28 * Convergence: true * √(Σ(yᵢ-ȳ)²)/n < 1.0e-08: true * Reached Maximum Number of Iterations: false * Objective Calls: 59

### What was nice about Julia?

Rating | Description |
---|---|

`Truncated(DN, lower, upper)` is a wonderfully simple way to define truncated distribution | |

A `logpdf` function that works for any distribution | |

Very clear and informative output |

### What was not so nice about Julia?

## R solution

R has a `truncnorm`

package for dealing with truncated normals.

odr=c(0.10,0.20,0.15,0.22,0.15,0.10,0.08,0.09,0.12) x = qnorm(odr) library(truncnorm) neglik_tn = function(x) { maxx = max(x) resfn = function(musigma) { -sum(log(dtruncnorm(x, a = -Inf, b= maxx, musigma[1], musigma[2]))) } resfn } neglikfn = neglik_tn(x) system.time(res <- optim(c(mean(x), sd(x)), neglikfn)) res

which gave the output

$par [1] -1.4294776 0.3555692 $value [1] -3.020876 $counts function gradient 57 NA $convergence [1] 0 $message NULL

### What was nice about R?

Rating | Description |
---|---|

Has a package for truncated normal | |

Gives me results right away; this wouldn't be such a plus if it weren't for Julia's slow compilation speed |

### What was not so nice about R?

Rating | Description |
---|---|

A couple of minor things: no log density for truncated normal; and no easy way to define truncated distribution for arbitrary distributions; sparse output (print) |

## Python solution

Even though I have no experience with Python, simple Google searches allowed me to come up with this solution. I have used the Anaconda distribution which saved me a lot of hassle in terms installing packages, as all the packages I need are pre-installed.

import numpy as np from scipy.optimize import minimize from scipy.stats import norm # generate the data odr=[0.10,0.20,0.15,0.22,0.15,0.10,0.08,0.09,0.12] Q_t = norm.ppf(odr) maxQ_t = max(Q_t) # define a function that will return a return to optimize based on the input data def neglik_trunc_tn(Q_t): maxQ_t = max(Q_t) def neglik_trunc_fn(musigma): return -sum(norm.logpdf(Q_t, musigma[0], musigma[1])) + len(Q_t)*norm.logcdf(maxQ_t, musigma[0], musigma[1]) return neglik_trunc_fn # the likelihood function to optimize neglik = neglik_trunc_tn(Q_t) # optimize! res = minimize(neglik, [np.mean(Q_t), np.std(Q_t)]) res

and these are the results

fun: -1.8930804441641282 hess_inv: array([[ 0.01759589, 0.00818596], [ 0.00818596, 0.00937868]]) jac: array([ -3.87430191e-07, 3.33786011e-06]) message: 'Optimization terminated successfully.' nfev: 40 nit: 6 njev: 10 status: 0 success: True x: array([-1.07334252, 0.25373624])

### What was nice about Python?

Rating | Description |
---|---|

Easy to google even for beginners | |

Gives me results right away |

### What was not so nice about Python?

Rating | Description |
---|---|

The output could be nicer, but a minor point |

## Conclusion

For such a simple optimization problem, you can't go wrong choosing any of the three languages/ecosystems. All three languages/ecosystems allow you to define closures which I find to be a good way to generate functions that embed the data that it needs, leaving you with just a function with the distribution parameters as the only arguments.

I would not choose to use Julia if I only need to run simple (sub-10 seconds) optimizations that I only intend to run once or twice due to long compilation times. If your problem is fairly standard you can get around this by using `PackageCompiler.jl`

to precompile the functions; but that feels like an over-complication for such a simple task that both R and Python/SciPy can handle easily.